- Abstract: This paper proposes a new strategy to train slice-level classifiers on CT based on the descriptors of the adjacent slices along the axis, each of which is extracted through a convolutional neural network (CNN). This method aims to predict the presence of ICH and classify it into 5 different sub-types. We exploit a two-stage training scheme. In the first stage, we treat a CT scan simply as a set of 2D images and train a state-of-the-art CNN classifier that was pretrained on ImageNet. During the training, each slice is sampled together with the 3 slices before and the 3 slices after it, which makes the batch size a multiple of 7. In the second stage, the output descriptors of each block of 7 consecutive slices obtained from stage 1 are stacked into an image and fed to another CNN for final prediction of the middle slice. Our model is entirely trained on the RSNA dataset and additionally evaluated on the CQ500 dataset, which adopts the same a set of labels but only on study level. We obtain a single model in the top 4% best-performing solutions of the RSNA ICH challenge, where model ensembles are allowed. Experiments also show that the proposed method significantly outperforms the baseline model on CQ500.
- Paper Type: both
- Track: short paper
- Keywords: CNNs, CT, Intracranial Hemorrhage Detection